Accurate structure prediction of

biomolecular interactions with AlphaFold 3

Josh Abramson, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel,

Olaf Ronneberger, Lindsay Willmore, Andrew J. Ballard, Joshua Bambrick,

Sebastian W. Bodenstein, David A. Evans, Chia-Chun Hung, Michael O’Neill, David Reiman,

Kathryn Tunyasuvunakool, Zachary Wu, Akvilė Žemgulytė, Eirini Arvaniti, Charles Beattie,

Ottavia Bertolli, Alex Bridgland, Alexey Cherepanov, Miles Congreve,

Alexander I. Cowen-Rivers, Andrew Cowie, Michael Figurnov, Fabian B. Fuchs,

Hannah Gladman, Rishub Jain, Yousuf A. Khan, Caroline M. R. Low, Kuba Perlin,

Anna Potapenko, Pascal Savy, Sukhdeep Singh, Adrian Stecula, Ashok Thillaisundaram,

Catherine Tong, Sergei Yakneen, Ellen D. Zhong, Michal Zielinski, Augustin Žídek,

Victor Bapst, Pushmeet Kohli, Max Jaderberg, Demis Hassabis & John M. Jumper

This is a PDF file of a peer-reviewed paper that has been accepted for publication.

Although unedited, the content has been subjected to preliminary formatting. Nature

is providing this early version of the typeset paper as a service to our authors and

readers. The text and figures will undergo copyediting and a proof review before the

paper is published in its final form. Please note that during the production process

errors may be discovered which could affect the content, and all legal disclaimers

apply.

Received: 19 December 2023

Accepted: 29 April 2024

Accelerated Article Preview

Published online xx xx xxxx

Cite this article as: Abramson, J. et al.

Accurate structure prediction of

biomolecular interactions with

AlphaFold 3. Nature https://doi.org/

10.1038/s41586-024-07487-w (2024)

https://doi.org/10.1038/s41586-024-07487-w

Nature | www.nature.com

Accelerated Article Preview

A C C E L E R A T E D A R TI C L E P R E VI E W

1 Accurate structure prediction of biomolecular

2 interactions with AlphaFold 3

3

Josh Abramson*1, Jonas Adler*1, Jack Dunger*1, Richard Evans*1, Tim Green*1 4 , Alexander

Pritzel*1, Olaf Ronneberger*1, Lindsay Willmore*1, Andrew J Ballard1

, Joshua Bambrick2 5 ,

Sebastian W Bodenstein1

, David A Evans1

, Chia-Chun Hung2

, Michael O'Neill1

, David Reiman1 6 ,

Kathryn Tunyasuvunakool1

, Zachary Wu1

, Akvilė Žemgulytė1

, Eirini Arvaniti3

, Charles Beattie3 7 ,

Ottavia Bertolli3

, Alex Bridgland3

, Alexey Cherepanov4

, Miles Congreve4 8 , Alexander I CowenRivers3

, Andrew Cowie3

, Michael Figurnov3

, Fabian B Fuchs3

, Hannah Gladman3

, Rishub Jain3 9 ,

Yousuf A Khan3

, Caroline M R Low4

, Kuba Perlin3

, Anna Potapenko3

, Pascal Savy4 10 , Sukhdeep

Singh3

, Adrian Stecula4

, Ashok Thillaisundaram3

, Catherine Tong4

, Sergei Yakneen4 11 , Ellen D

Zhong3

, Michal Zielinski3

, Augustin Žídek3

, Victor Bapst†1, Pushmeet Kohli†1, Max Jaderberg†2 12 ,

Demis Hassabis†1,2, John M Jumper†1 13

14

15

* 16 Contributed equally

1 17 Core Contributor, Google DeepMind, London, UK

2 18 Core Contributor, Isomorphic Labs, London, UK

3 19 Google DeepMind, London, UK

4 20 Isomorphic Labs, London, UK

† 21 Jointly supervised

22

23 Corresponding author emails:

24 J. J. - jumper@google.com; D.H. - dhcontact@google.com; M.J. - jaderberg@isomorphiclabs.com

25

26

The introduction of AlphaFold 2

1 27 has spurred a revolution in modelling the structure of

28 proteins and their interactions, enabling a huge range of applications in protein modelling

and design2–6 29 . In this paper, we describe our AlphaFold 3 model with a substantially

30 updated diffusion-based architecture, which is capable of joint structure prediction of

31 complexes including proteins, nucleic acids, small molecules, ions, and modified residues.

32 The new AlphaFold model demonstrates significantly improved accuracy over many

33 previous specialised tools: far greater accuracy on protein-ligand interactions than state of

34 the art docking tools, much higher accuracy on protein-nucleic acid interactions than

35 nucleic-acid-specific predictors, and significantly higher antibody-antigen prediction

accuracy than AlphaFold-Multimer v2.37,8 36 . Together these results show that high accuracy

37 modelling across biomolecular space is possible within a single unified deep learning

ACCELERATED

38 framework.

ARTICLE PREVIEW

39 Main Text

40 Introduction

41 Accurate models of biological complexes are critical to our understanding of cellular functions

and for the rational design of therapeutics2–4,9 42 . Enormous progress has been achieved in protein

structure prediction with the development of AlphaFold1 43 , and the field has grown tremendously

with a number of later methods that build on the ideas and techniques of AlphaFold 210–12 44 .

45 Almost immediately after AlphaFold became available, it was shown that simple input

modifications would enable surprisingly accurate protein interaction predictions13–15 46 and that

47 training AlphaFold 2 specifically for protein interaction prediction yielded a highly accurate

system7 48 .

49

50 These successes lead to the question of whether it is possible to accurately predict the structure

51 of complexes containing a much wider range of biomolecules, including ligands, ions, nucleic

52 acids, and modified residues, within a deep learning framework. A wide range of predictors for

various specific interaction types have been developed16–28 53 , as well as one generalist method

developed concurrently with the present work29 54 , but the accuracy of such deep learning attempts

has been mixed and often below that of physics-inspired methods30,31 55 . Almost all these methods

56 are also highly specialised to particular interaction types and cannot predict the structure of

57 general biomolecular complexes containing many types of entities.

58

59 Here, we present AlphaFold 3 (AF3), a model that is capable of high accuracy prediction of

complexes containing nearly all molecular types present in the Protein Data Bank32 60 (PDB) (Fig.

61 1a,b). In all but one category it achieves a significantly higher performance than strong methods

62 that specialise in just the given task (Fig. 1c, Extended Data Table 1) including higher accuracy

63 at protein structure and the structure of protein-protein interactions.

64

65 This is achieved by a substantial evolution of the AlphaFold 2 architecture and training

66 procedure (Fig. 1d) both to accommodate more general chemical structures and to improve the

67 data efficiency of learning. The system reduces the amount of multiple sequence alignment

68 (MSA) processing by replacing the AlphaFold 2 Evoformer with the simpler Pairformer Module

69 (Fig. 2a). Furthermore it directly predicts the raw atom coordinates with a Diffusion Module,

70 replacing the AlphaFold 2 Structure Module that operated on amino-acid-specific frames and

71 side chain torsion angles (Fig. 2b). The multiscale nature of the diffusion process (low noise

72 levels induce the network to improve local structure) also allow us to eliminate stereochemical

73 losses and most special handling of bonding patterns in the network, easily accommodating

74 arbitrary chemical components.

75

ACCELERATED ARTICLE PREVIEW

76 Network architecture and training

77 The overall structure of AF3 (Fig. 1d, Supplementary Methods 3) echoes that of AlphaFold 2

78 with a large trunk evolving a pairwise representation of the chemical complex followed by a

79 Structure Module that uses the pairwise representation to generate explicit atomic positions, but

80 there are large differences in each major component. These modifications were driven both by

81 the need to accommodate a wide range of chemical entities without excessive special-casing and

82 by observations of AlphaFold 2 performance with different modifications. Within the trunk,

83 MSA processing is substantially de-emphasized with a much smaller and simpler MSA

84 embedding block (Supplementary Methods 3.3). Compared to the original Evoformer from

85 AlphaFold 2 the number of blocks are reduced to four, the processing of the MSA representation

86 uses an inexpensive pair-weighted averaging, and only the pair representation is used for later

87 processing steps. The \"Pairformer\" (Fig. 2a, Supplementary Methods 3.6) replaces the

88 \"Evoformer\" of AlphaFold 2 as the dominant processing block. It operates only on the pair

89 representation and the single representation; the MSA representation is not retained and all

90 information passes via the pair representation. The pair processing and the number of blocks (48)

91 is largely unchanged from AlphaFold 2. The resulting pair and single representation together

92 with the input representation are passed to the new Diffusion Module (Fig. 2b) that replaces the

93 Structure Module of AlphaFold 2.

94

95 The Diffusion Module (Fig. 2b, Supplementary Methods 3.7) operates directly on raw atom

96 coordinates, and on a coarse abstract token representation, without rotational frames or any

97 equivariant processing. We had observed in AlphaFold 2 that removing most of the complexity

98 of the Structure Module had only a modest effect on prediction accuracy, and maintaining the

99 backbone frame and side chain torsion representation add quite a bit of complexity for general

100 molecular graphs. Similarly AlphaFold 2 required carefully tuned stereochemical violation

101 penalties during training to enforce chemical plausibility of the resulting structures. We use a

relatively standard diffusion approach34 102 in which the diffusion model is trained to receive

103 “noised” atomic coordinates then predict the true coordinates. This task requires the network to

104 learn protein structure at a variety of length scales, where the denoising task at small noise

105 emphasises understanding very local stereochemistry and the denoising task at high noise

106 emphasises large-scale structure of the system. At inference time, random noise is sampled and

107 then recurrently denoised to produce a final structure. Importantly, this is a generative training

108 procedure which produces a distribution of answers. This means that, for each answer, the local

109 structure will be sharply defined (e.g. side chain bond geometry) even when the network is

110 uncertain about the positions. For this reason, we are able to avoid both torsion-based

111 parametrizations of the residues and violation losses on the structure, while handling the full

complexity of general ligands. Similarly to some recent work35 112 , we find that no invariance or

113 equivariance with respect to global rotations and translation of the molecule are required in the

114 architecture and so we omit them to simplify the machine learning architecture.

115

116 The use of a generative diffusion approach comes with some technical challenges that we needed

to address. The biggest issue is that generative models are prone to hallucination36 117 where the

ACCELERATED ARTICLE PREVIEW

118 model may invent plausible-looking structure even in unstructured regions. To counteract this

119 effect, we use a novel cross-distillation method where we enrich the training data with

AlphaFold-Multimer v2.37,8 120 predicted structures. In these structures, unstructured regions are

121 typically represented by long extended loops instead of compact structures and training on them

122 “teaches” AlphaFold 3 to mimic this behaviour. This cross-distillation greatly reduced the

123 hallucination behaviour of AF3 (Extended Data Fig. 1 for disorder prediction results on the

CAID 237 124 benchmark set).

125

126 We also developed confidence measures that predict the atom-level and pairwise errors in our

127 final structures. In AlphaFold 2, this was done directly by regressing the error in the output of the

128 Structure Module during training. This procedure is not applicable to diffusion training however,

129 since only a single step of the diffusion is trained instead of a full structure generation (Fig. 2c).

130 To remedy this, we developed a diffusion “rollout” procedure for the full structure prediction

131 generation during training (using a larger step size than normal; see Fig. 2c \"mini-rollout\"). This

132 predicted structure is then used to permute the symmetric ground truth chains and ligands, and to

133 compute the performance metrics to train the confidence head. The confidence head uses the

134 pairwise representation to predict the LDDT (pLDDT) and a predicted aligned error (PAE)

135 matrix as in AlphaFold 2, as well as a distance error matrix (PDE) which is the error in the

136 distance matrix of the predicted structure as compared to the true structure (see Supplementary

137 Methods 4.3 for details).

138

139 Fig. 2d shows that during initial training the model learns quickly to predict the local structures

140 (all intra chain metrics go up quickly and reach 97% of the maximum performance within the

141 first 20k training steps) while the model needs considerably longer to learn the global

142 constellation (the interface metrics go up slowly and protein-protein interface LDDT passes the

143 97% bar only after 60k steps). During AF3 development we observed that some model

144 capabilities topped out relatively early and started to decline (most likely due to overfitting to the

145 limited number of training samples for this capability) while other capabilities were still

146 undertrained. We addressed this by increasing / decreasing the sampling probability for the

147 corresponding training sets (Supplementary Methods 2.5.1) and by an early stopping using a

148 weighted average of all above metrics and some additional metrics to select the best model

149 checkpoint (Supplementary Table 7). The fine tuning stages with the larger crop sizes improve

150 the model on all metrics with an especially high uplift on protein-protein interfaces (Extended

151 Data Fig. 2).

152 Accuracy across complex types

153 AF3 can predict structures from input polymer sequences, residue modifications, and ligand

154 SMILES. In Fig. 3 we show a selection of examples highlighting the ability of the model to

155 generalise to a number of biologically important and therapeutically relevant modalities. In

156 selecting these examples, we considered novelty in terms of the similarity of individual chains

157 and interfaces to the training set (additional information in Supplementary Methods 8.1).

158

ACCELERATED ARTICLE PREVIEW

159 We evaluate performance of the system on recent interface-specific benchmarks for each

160 complex type (Fig. 1c, Extended Data Table 1). Performance on protein-ligand interfaces was

161 evaluated on the PoseBusters benchmark set, composed of 428 protein-ligand structures released

162 to the PDB in 2021 or later. Since our standard training cutoff date is in 2021, we trained a

163 separate AF3 model with an earlier training set cutoff (see Methods for details). Accuracy on the

164 PoseBusters set is reported as the percentage of protein-ligand pairs with pocket-aligned ligand

165 RMSD of less than 2 Å. The baseline models come in two categories: those that use only protein

166 sequence and ligand SMILES as input and those that additionally leak information from the

167 solved protein-ligand test structure. Traditional docking methods use the latter privileged

168 information, even though that information would not be available in real world use cases. Even

so, AlphaFold 3 greatly outperforms classical docking tools like Vina38,39 169 even while not using

any structural inputs (Fisher exact p=2.27 * 10-13 170 ) and greatly outperforms all other true blind

docking like RoseTTAFold All-Atom (p=4.45 * 10-25 171 ). Extended Data Fig. 3 shows three

172 examples where AlphaFold 3 achieves accurate predictions but docking tools Vina and Gold do

not38 173 . PoseBusters analysis was done using a 2019-09-30 training cutoff for AlphaFold 3 to

174 ensure the model was not trained on any PoseBusters structures. To compare to RoseTTAFold

175 All-Atom results, we used PoseBusters Version 1. Version 2 (crystal contacts removed from the

176 benchmark set) results including quality metrics are shown in Extended Data Fig. 4b-f and in

177 Extended Data Table 1. We use multiple seeds to ensure correct chirality and avoid slight

178 protein-ligand clashing (as opposed to a method like diffusion guidance to enforce) but are

179 typically able to produce high quality stereochemistry. Separately, we also train a version of

180 AlphaFold 3 that receives the “pocket information” as used in some recent deep learning

work24,26 181 (Extended Data Fig. 4a for results).

182

183 AF3 predicts protein-nucleic complexes and RNA structures with higher accuracy than

RoseTTAFold2NA40 184 (Fig. 1c second plot). As RoseTTAFold2NA is only validated on structures

185 below 1000 residues, we use only structures below 1000 residues from our Recent PDB

186 evaluation set for this comparison (see Methods for details). AlphaFold 3 is able to predict

187 protein-nucleic structures with thousands of residues, an example of which is shown in Fig. 3a.

188 Note that we do not compare directly to RoseTTAFold All-Atom, but benchmarks indicate that

189 RoseTTAFold All-Atom is comparable to slightly less accurate than RoseTTAFold2NA for

nucleic acid predictions29 190 .

191

192 We also evaluated AF3 performance on the 10 publicly available CASP15 RNA targets: We

achieve a higher average performance than RoseTTAFold2NA and AIchemy_RNA27 193 (the best

AI-based submission in CASP1518,31 194 ) on the respective common subsets of our and their

195 predictions (see Extended Data Fig. 5a for detailed results). We do not reach the performance

of the best human-expert-aided CASP15 submission AIchemy_RNA241 196 (Fig. 1c, centre left).

197 Due to limited dataset sizes, we do not report significance test statistics here. Further analysis of

198 the accuracy of predicting nucleic acids alone (without proteins) is shown in Extended Data

199 Fig. 5b.

200

ACCELERATED ARTICLE PREVIEW

201 Covalent modifications (bonded ligands, glycosylation, and modified protein residues and

202 nucleic acid bases) are also accurately predicted by AF3 (Fig. 1c, centre right). Modifications

203 include those to any polymer residue (protein, RNA or DNA). We report accuracy as the

204 percentage of successful predictions (pocket RMSD < 2 Å). We apply quality-filters to the

205 bonded ligands and glycosylation dataset (as does PoseBusters): We only include ligands with

206 high-quality experimental data (ranking_model_fit > 0.5 according to the RCSB structure

207 validation report, that is, X-ray structures with a model quality above the median). As with the

208 PoseBusters set, the bonded ligands and glycosylation datasets are not filtered by homology to

209 the training data set. Filtering based on the bound polymer chain homology (using polymer

210 template similarity < 40) yielded only 5 clusters for bonded ligands and 7 clusters for

211 glycosylation. We exclude multi-residue glycans here, because the RCSB validation report does

212 not provide a ranking_model_fit value for them. The percentage of successful predictions

213 (pocket RMSD < 2 Å) for multi-residue glycans on all-quality experimental data is 42.1%

214 (N=131 clusters) which is slightly lower than the success rate for single-residue glycans on all215 quality experimental data of 46.1% (N=167). The modified residues dataset is filtered similarly

216 to our other polymer test sets: it contains only modified residues in polymer chains with low

217 homology to the training set (see Methods for details). See Extended Data Table 1 for detailed

218 results; Extended Data Fig. 6 for examples of predicted protein, DNA, and RNA structures with

219 covalent modifications including analysis of the impact phosphorylation has on predictions.

220

221 While expanding in modelling capabilities, AF3 has also improved in protein complex accuracy

relative to AlphaFold-Multimer v2.37,8 222 (AF-M 2.3). Generally, protein-protein prediction success

(DockQ42 > 0.23) has increased (paired Wilcoxon signed-rank test, p=1.8 * 10-18 223 ), with antibody224 protein interaction prediction in particular showing a marked improvement (Fig. 1c right, paired

Wilcoxon signed-rank test, p=6.5 * 10-5 225 , predictions top ranked from 1000 rather than the

226 typical 5 seeds, see Fig. 5a for details). Protein monomer LDDT improvement is also significant

(paired Wilcoxon signed-rank test, p=1.7 * 10-34 227 ). AF3 has a very similar dependence on MSA

228 depth to AF-M 2.3; proteins with shallow MSAs are predicted with lower accuracy (see

229 Extended Data Fig. 7a for a comparison of the dependence of single chain LDDT on MSA

230 depth).

231

232 Predicted confidences track accuracy

233 As with AlphaFold 2, AlphaFold 3 confidence measures are well-calibrated with accuracy. Our

234 confidence analysis is performed on the recent PDB evaluation set, with no homology filtering

235 and including peptides. The ligands category is filtered to high quality experimental structures as

236 described above, and considers standard non-bonded ligands only. See Extended Data Fig. 8 for

237 a similar assessment on bonded ligand and other interfaces. All statistics are cluster-weighted

238 (see Methods for details) and consider the top-ranked prediction only (see Supplementary

239 Methods 5.9.3 for ranking details).

240

ACCELERATED ARTICLE PREVIEW

In Fig. 4a top row we plot chain pair ipTM (interface predicted TM score43 241 ; see Supplementary

242 Methods 5.9.1) against interface accuracy measures: protein-protein DockQ, protein-nucleic

243 iLDDT, and protein-ligand success, with success defined as percent of examples under

244 thresholded pocked-aligned RMSD values. In Fig. 4a bottom row we plot average pLDDT per

245 protein, nucleotide or ligand entity against our bespoke LDDT_to_polymer metric (for metrics

246 details, see Methods), which is closely related to the training target of the pLDDT predictor.

247

248 In Fig. 4b-e we highlight a single example prediction of 7T82, where per-atom pLDDT

249 colouring identifies unconfident chain tails, somewhat confident interfaces, and otherwise

250 confident secondary structure. In Fig. 4c the same prediction is coloured by chain, along with

251 DockQ interface scores in Fig. 4d and per-chain colouring displayed on the axes for reference.

252 We see from Fig. 4e that PAE confidence is high for pink-grey and blue-orange residue pairs

253 where DockQ > 0.7, and least confident about pink-orange and pink-blue residue pairs which

254 have DockQ ≈ 0. See Extended Data Fig. 5c-d for a similar PAE analysis on an example with

255 protein and nucleic acid chains.

256

257 Model limitations

258 We note model limitations of AlphaFold 3 with respect to stereochemistry, hallucinations,

259 dynamics, and accuracy for certain targets.

260

261 On stereochemistry, we note two main classes of violations. The first is that the model outputs do

262 not always respect chirality (Fig. 5b), despite the model receiving reference structures with

263 correct chirality as input features. To address this in the PoseBusters benchmark, we included a

264 penalty for chirality violation in our ranking formula for model predictions. Despite this, we still

265 observe a chirality violation rate of 4.4% in the benchmark. The second class of stereochemical

266 violations is a tendency of the model to occasionally produce overlapping (“clashing”) atoms in

267 the predictions. This sometimes manifests as extreme violations in homomers where entire

268 chains have been observed to overlap (Fig. 5e). Penalising clashes during ranking (see

269 Supplementary Methods 5.9.3) reduces the occurrence of this failure mode but does not

270 eliminate them. Almost all remaining clashes occur for protein-nucleic complexes with both

271 greater than 100 nucleotides and greater than 2,000 residues in total.

272

273 We note that the switch from the non-generative AlphaFold 2 model to the diffusion-based

274 AlphaFold 3 model introduces the challenge of spurious structural order (hallucinations) in

275 disordered regions (Fig. 5d, Extended Data Fig. 1). While hallucinated regions are typically

276 marked as very low confidence, they can lack the distinctive ribbon-like appearance that

277 AlphaFold 2 produces in disordered regions. To encourage ribbon-like predictions in AF3, we

278 use distillation training from AlphaFold 2 predictions, and we add a ranking term to encourage

results with more solvent accessible surface area37 279 .

280

ACCELERATED ARTICLE PREVIEW

281 A key limitation of protein structure prediction models is that they typically predict static

282 structures as seen in the PDB, not the dynamical behaviour of biomolecular systems in solution.

283 This limitation persists for AlphaFold 3, where multiple random seeds for either the diffusion

284 head or the overall network do not produce an approximation of the solution ensemble.

285

286 In some cases, the modelled conformational state may not be correct or comprehensive given the

287 specified ligands and other inputs. As an example, E3 ubiquitin ligases natively adopt an open

288 conformation in an apo state and have only been observed in a closed state when bound to

ligands, but AF3 exclusively predicts the closed state for both holo and apo systems44 289 (Fig. 5c).

290 Many methods have been developed, particularly around MSA resampling, which assist in

generating diversity from previous AlphaFold models45–47 291 and may also assist in multi-state

292 prediction with AF3.

293

294 Despite the large advance in modelling accuracy in AlphaFold 3, there are still many targets for

295 which accurate modelling can be challenging. To obtain the highest accuracy, it may be

296 necessary to generate a large number of predictions and rank them, which incurs an extra

297 computational cost. A class of targets where we observe this effect strongly is antibody-antigen

complexes similar to other recent work48 298 . Fig. 5a shows that for AlphaFold 3, top-ranked

299 predictions keep improving with more model seeds, even at as many as 1000 (Wilcoxon signed

rank test between 5 and 1000 seeds, p=2.0 * 10-5 300 for % correct and p=0.009 for % very high

301 accuracy; ranking by protein-protein interface ipTM). This large improvement with many seeds

302 isn’t observed in general for other classes of molecules (see Extended Data Fig. 7b). Using only

303 one diffusion sample per model seed for the AF3 predictions rather than five (not illustrated)

304 does not change results significantly, indicating that running more model seeds is necessary for

305 antibody score improvements, rather than just more diffusion samples.

306

307 Discussion

308 The core challenge of molecular biology is to understand and ultimately regulate the complex

309 atomic interactions of biological systems. The AlphaFold 3 model takes a large step in this

310 direction, demonstrating that it is possible to accurately predict the structure of a wide range of

311 biomolecular systems in a unified framework. While there are still substantial challenges to

312 achieve highly accurate predictions across all interaction types, we demonstrate that it is possible

313 to build a deep learning system that shows strong coverage and generalisation for all these

314 interactions. We also demonstrate that the lack of cross-entity evolutionary information is not a

315 substantial blocker to progress in predicting these interactions, and moreover substantial

316 improvement in antibody results suggests AlphaFold-derived methods are able to model the

317 chemistry and physics of classes of molecular interactions without dependence on MSAs.

318 Finally, the large improvement in protein-ligand structure prediction shows that it is possible to

319 handle the wide diversity of chemical space within a general deep learning framework and

320 without resorting to an artificial separation between protein structure prediction and ligand

321 docking.

ACCELERATED ARTICLE PREVIEW

322

323 The development of bottom-up modelling of cellular components is a key step in unravelling the

324 complexity of molecular regulation within the cell, and the performance of AlphaFold 3 shows

325 that developing the right deep learning frameworks can massively reduce the amount of data

326 required to obtain biologically relevant performance on these tasks and amplify the impact of the

327 data already collected. We expect that structural modelling will continue to improve not only due

328 to deep learning advances but also because continuing methodological advances in experimental

329 structure determination, such as the dramatic improvements in cryo electron microscopy and

330 tomography, will provide a wealth of new training data to further the improve the generalisation

331 capability of such models. The parallel developments of experimental and computational

332 methods promise to propel us further into an era of structurally informed biological

333 understanding and therapeutic development.

334 Figure Captions

335

336 Fig. 1 | AlphaFold 3 accurately predicts structures across biomolecular complexes. a, Example structure

337 predicted with AF3: bacterial CRP/FNR family transcriptional regulator protein bound to DNA and cGMP (PDB ID

7PZB, full complex LDDT33 338 : 82.8, GDT: 90.1). b, Example structure predicted with AF3: human coronavirus OC43

339 spike protein, 4665 residues, heavily glycosylated and bound by neutralising antibodies (PDB ID 7PNM, full

340 complex LDDT: 83.0, GDT: 83.1). c, Performance on PoseBusters (V1, August 2023 release), our Recent PDB

341 evaluation set, and CASP15 RNA. Metrics are % of pocket-aligned ligand RMSD < 2 Å for ligands and covalent

342 modifications, interface LDDT for protein-nucleic acid complexes, LDDT for nucleic acid and protein monomers,

343 and % DockQ > 0.23 for protein-protein and protein-antibody interfaces. All scores are reported from the top

344 confidence-ranked sample out of 5 model seeds (each with 5 diffusion samples), except for protein-antibody scores

345 which were ranked across 1000 model seeds for both models (each AF3 seed with 5 diffusion samples). See

346 Methods for sampling and ranking details. For ligands, N indicates number of targets; for nucleic acids, N indicates

347 number of structures; for modifications, N indicates clusters, and for proteins N indicates clusters. Bar heights

348 indicate means; error bars indicate exact binomial distribution 95% confidence intervals for PoseBusters and via

349 10,000 bootstrap resamples for all others. Significance levels calculated via two-sided Fisher’s Exact Test for

350 PoseBusters and via two-sided Wilcoxon signed rank test for all others. *** for p < 0.001, ** for p<0.01. P-values

(left to right): 2.27*10-13, 2.57*10-3

, 2.78*10-3

, 7.28*10-12, 1.81*10-18, 6.54*10-5

, and 1.74*10-34 351 . d, AF3 architecture

352 for inference. Rectangles represent processing modules, arrows show the data flow. yellow: input data, blue: abstract

353 network activations, green: output data. Coloured balls represent physical atom coordinates.

354

355 Fig. 2 | Architectural and training details. a, Pairformer Module. Input and output: pair representation with

356 dimension (n, n, c) and single representation with dimension (n, c). n: number of tokens (polymer residues and

357 atoms), c: number of channels (128 for the pair representation, 384 for the single representation). Each of the 48

358 blocks has an independent set of trainable parameters. b, Diffusion Module. Input: coarse arrays depict per-token

359 representations (green: inputs, blue: pair, red: single). Fine arrays depict per-atom representations. Coloured balls

360 represent physical atom coordinates. c, training setup (distogram head omitted) starting from the end of the network

361 trunk. Coloured arrays: activations from the network trunk (green: inputs, blue: pair, red: single). Blue arrows:

362 abstract activation arrays; yellow arrows: ground truth data; green arrows: predicted data. Stop sign: stop gradient

363 operation. Both depicted Diffusion Modules share weights. d, Training curves for initial training and fine tuning

364 stages, showing LDDT on our evaluation set as a function of optimizer steps. The scatter plot shows the raw data

365 points and the lines show the smoothed performance using a median filter with a kernel width of 9 data points. The

366 crosses mark the point where the smoothed performance reaches 97% of its initial training maximum.

ACCELERATED ARTICLE PREVIEW

367

368 Fig. 3 | Examples of predicted complexes. : Selected structure predictions from AF3. Predicted protein chains are

369 shown in blue (predicted antibody in green), predicted ligands and glycans in orange, predicted RNA in purple, and

370 ground truth in grey. a, Human 40S small ribosomal subunit (7663 residues) including 18S ribosomal RNA and

Met-tRNAi

Met 371 (opaque purple) in complex with translation initiation factors eIF1A and eIF5B (opaque blue; PDB ID

372 7TQL, full complex LDDT: 87.7, GDT: 86.9). b,Glycosylated globular portion of an EXTL3 homodimer (PDB ID

373 7AU2, mean pocked-aligned RMSD: 1.10 Å). c, Mesothelin C-terminal peptide bound to the monoclonal antibody

374 15B6 (PDB ID 7U8C, DockQ: 0.85). d, LGK974, a clinical stage inhibitor, bound to PORCN in complex with the

375 WNT3A peptide (PDB ID 7URD, ligand RMSD 1.00 Å). e, (5S,6S)-O7-sulfo DADH bound to the AziU3/U2

376 complex with a novel fold (PDB ID 7WUX, ligand RMSD 1.92 Å). f, Analog of NIH-12848 bound to an allosteric

377 site of PI5P4Kγ (PDB ID 7QIE, ligand RMSD 0.37 Å).

378

379 Fig. 4 | AlphaFold 3 confidences track accuracy. a (top row), Accuracy of protein-containing interfaces as a

380 function of chain pair ipTM. a (bottom row), LDDT_to_polymer accuracy evaluated for various chain types as a

381 function of chain-averaged pLDDT. Box, centerline, and whiskers boundaries are at (25%, 75%) intervals, median,

382 and (5%, 95%) intervals. N values report the number of clusters in each band. b, Predicted structure of PDB ID

383 7T82 coloured by pLDDT (orange: 0-50, yellow: 50-70, cyan 70-90, and blue 90-100). c, same prediction coloured

384 by chain. d, DockQ scores for protein-protein interfaces. e, Predicted Aligned Error (PAE) matrix of same

385 prediction (darker is more confident), with chain colouring of panel c on side-bars. Dashed black lines indicate chain

386 boundaries.

387

388 Fig. 5 | Model limitations. a, Antibody prediction quality increases with the number of model seeds. Quality of top389 ranked, low homology, antibody-antigen interface predictions as a function of number of seeds. Each datapoint

390 shows the mean over 1,000 random samples (with replacement) of seeds to rank over, out of 1200 seeds. Confidence

391 intervals are 95% bootstraps over 10,000 resamples of cluster scores at each datapoint. Samples per interface ranked

392 by protein-protein ipTM. Significance tests are by a two-sided Wilcoxon signed rank test. N = 65 clusters. *** for p

< 0.001. P-values: 2.0 * 10-5 393 for % correct and p=0.009 for % very high accuracy. b, Prediction (coloured) and

394 ground truth (grey) structures of Thermotoga maritima alpha-glucuronidase and beta-D-glucuronic acid, a target

395 from the PoseBusters set (PDB ID 7CTM). AF3 predicts alpha-D-glucuronic acid, differing chiral centre indicated

396 by an asterisk. The prediction shown is top-ranked by ligand-protein ipTM and with a chirality and clash penalty. c,

397 Conformation coverage is limited. Ground truth structures (grey) of cereblon in open (apo PDB ID 8CVP, left) and

398 closed (holo mezigdomide-bound, PDB ID 8D7U, right) conformations. Predictions (blue) of both apo (with 10

399 overlaid samples) and holo structures are in the closed conformation. Dashed line indicates distance between the N400 terminal Lon protease-like and C-terminal thalidomide-binding domain. d, A nuclear pore complex with 1,854

401 unresolved residues (PDB ID 7F60). Ground truth (left) and predictions from AF-M 2.3 (middle) and AF3 (right). e,

402 Prediction of a trinucleosome with overlapping DNA (pink) and protein (blue) chains (PDB ID 7PEU); highlighted

403 are overlapping protein chains B and J and self-overlapping DNA chain AA. Unless otherwise stated, predictions are

404 top-ranked by our global complex ranking metric with chiral mismatch and steric clash penalties (see

405 Supplementary Methods 5.9.1).

406

407 Citations

408 1. Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–

409 589 (2021).

410 2. Kreitz, J. et al. Programmable protein delivery with a bacterial contractile injection system.

411 Nature 616, 357–364 (2023).

ACCELERATED ARTICLE PREVIEW

412 3. Lim, Y. et al. In silico protein interaction screening uncovers DONSON’s role in replication

413 initiation. Science 381, eadi3448 (2023).

414 4. Mosalaganti, S. et al. AI-based structure prediction empowers integrative structural analysis of

415 human nuclear pores. Science 376, eabm9506 (2022).

416 5. Anand, N. & Achim, T. Protein Structure and Sequence Generation with Equivariant Denoising

417 Diffusion Probabilistic Models. arXiv preprint arXiv:2205.15019 (2022)

418 doi:10.48550/arXiv.2205.15019.

419 6. Yang, Z., Zeng, X., Zhao, Y. & Chen, R. AlphaFold2 and its applications in the fields of biology

420 and medicine. Signal Transduction and Targeted Therapy 8, 1–14 (2023).

421 7. Evans, R. et al. Protein complex prediction with AlphaFold-Multimer. bioRxiv

422 2021.10.04.463034 (2022) doi:10.1101/2021.10.04.463034.

423 8. alphafold/docs/technical_note_v2.3.0.md at main · google-deepmind/alphafold. GitHub

424 https://github.com/google-deepmind/alphafold/blob/main/docs/technical_note_v2.3.0.md.

425 9. Structure-based drug design with geometric deep learning. Curr. Opin. Struct. Biol. 79, 102548

426 (2023).

427 10.Lin, Z. et al. Evolutionary-scale prediction of atomic-level protein structure with a language

428 model. Science 379, 1123–1130 (2023).

429 11.Baek, M. et al. Accurate prediction of protein structures and interactions using a three-track

430 neural network. Science (2021) doi:10.1126/science.abj8754.

431 12.Wu, R. et al. High-resolution de novo structure prediction from primary sequence. bioRxiv

432 2022.07.21.500999 (2022) doi:10.1101/2022.07.21.500999.

433 13.Bryant, P., Pozzati, G. & Elofsson, A. Improved prediction of protein-protein interactions using

434 AlphaFold2. Nat. Commun. 13, 1–11 (2022).

435 14.[No title]. X (formerly Twitter)

436 https://twitter.com/Ag_smith/status/1417063635000598528?lang=en-GB.

437 15.[No title]. X (formerly Twitter)

438 https://twitter.com/minkbaek/status/1417538291709071362?lang=en.

ACCELERATED ARTICLE PREVIEW

439 16.Qiao, Z., Nie, W., Vahdat, A., Miller, T. F., III & Anandkumar, A. State-specific protein-ligand

440 complex structure prediction with a multi-scale deep generative model. arXiv preprint

441 arXiv:2209.15171 (2022) doi:10.48550/arXiv.2209.15171.

442 17.Nakata, S., Mori, Y. & Tanaka, S. End-to-end protein–ligand complex structure generation with

443 diffusion-based generative models. BMC Bioinformatics 24, 1–18 (2023).

444 18.Baek, M. et al. Accurate prediction of protein–nucleic acid complexes using RoseTTAFoldNA.

445 Nat. Methods 1–5 (2023).

446 19.Townshend, R. J. L. et al. Geometric deep learning of RNA structure. Science 373, 1047–1051

447 (2021).

448 20.Jiang, D. et al. InteractionGraphNet: A Novel and Efficient Deep Graph Representation Learning

449 Framework for Accurate Protein-Ligand Interaction Predictions. J. Med. Chem. 64, (2021).

450 21.Jiang, H. et al. Predicting Protein–Ligand Docking Structure with Graph Neural Network. J.

451 Chem. Inf. Model. (2022) doi:10.1021/acs.jcim.2c00127.

452 22.Corso, G., Stärk, H., Jing, B., Barzilay, R. & Jaakkola, T. DiffDock: Diffusion Steps, Twists, and

453 Turns for Molecular Docking. arXiv preprint arXiv:2210.01776 (2022)

454 doi:10.48550/arXiv.2210.01776.

455 23.Stärk, H., Ganea, O.-E., Pattanaik, L., Barzilay, R. & Jaakkola, T. EquiBind: Geometric Deep

456 Learning for Drug Binding Structure Prediction. arXiv preprint arXiv:2202.05146 (2022)

457 doi:10.48550/arXiv.2202.05146.

458 24.Liao, Z. et al. DeepDock: Enhancing Ligand-protein Interaction Prediction by a Combination of

459 Ligand and Structure Information. 2019 IEEE International Conference on Bioinformatics and

460 Biomedicine (BIBM), San Diego, CA, USA, 311 – 317 (2019).

461 25.Lu, W. et al. TANKBind: Trigonometry-Aware Neural NetworKs for Drug-Protein Binding

462 Structure Prediction. Adv. Neural Inf. Process. Syst. 35, 7236–7249 (2022).

463 26.Zhou, G. et al. Uni-Mol: A Universal 3D Molecular Representation Learning Framework. (2023)

464 doi:10.26434/chemrxiv-2022-jjm0j-v4.

465 27.Shen, T. et al. E2Efold-3D: End-to-End Deep Learning Method for accurate de novo RNA 3D

ACCELERATED ARTICLE PREVIEW

466 Structure Prediction. (2022).

467 28.van Dijk, M. & Bonvin, A. M. J. J. Pushing the limits of what is achievable in protein–DNA

468 docking: benchmarking HADDOCK’s performance. Nucleic Acids Res. 38, 5634–5647 (2010).

469 29.Krishna, R. et al. Generalized biomolecular modeling and design with RoseTTAFold All-Atom.

470 Science 384, eadl2528 (2024).

471 30.Buttenschoen, M., Morris, G. M. & Deane, C. M. PoseBusters: AI-based docking methods fail to

472 generate physically valid poses or generalise to novel sequences. arXiv preprint arXiv:2308.05777

473 (2023) doi:10.48550/arXiv.2308.05777.

474 31.Das, R. et al. Assessment of three-dimensional RNA structure prediction in CASP15. Proteins

475 91, (2023).

476 32.Berman, H. M. et al. The Protein Data Bank. Nucleic Acids Res. 28, 235–242 (2000).

477 33.Mariani, V., Biasini, M., Barbato, A. & Schwede, T. lDDT: a local superposition-free score for

478 comparing protein structures and models using distance difference tests. Bioinformatics 29, 2722–

479 2728 (2013).

480 34.Karras, T., Aittala, M., Aila, T. & Laine, S. Elucidating the Design Space of Diffusion-Based

481 Generative Models. arXiv preprint arXiv:2206.00364 (2022) doi:10.48550/arXiv.2206.00364.

482 35.Wang, Y., Elhag, A. A., Jaitly, N., Susskind, J. M. & Bautista, M. A. Generating Molecular

483 Conformer Fields. arXiv preprint arXiv:2311.17932 (2023) doi:10.48550/arXiv.2311.17932.

484 36.Ji, Z. et al. Survey of Hallucination in Natural Language Generation. arXiv preprint

485 arXiv:2202.03629 (2022) doi:10.1145/3571730.

486 37.Del Conte, A. et al. Critical assessment of protein intrinsic disorder prediction (CAID) - Results

487 of round 2. Proteins: Struct. Funct. Bioinf. 91, 1925–1934 (2023).

488 38.Trott, O. & Olson, A. J. AutoDock Vina: Improving the speed and accuracy of docking with a

489 new scoring function, efficient optimization, and multithreading. J. Comput. Chem. 31, 455–461

490 (2010).

491 39.Miller, E. B. et al. Reliable and Accurate Solution to the Induced Fit Docking Problem for

492 Protein–Ligand Binding. J. Chem. Theory Comput. (2021) doi:10.1021/acs.jctc.1c00136.

ACCELERATED ARTICLE PREVIEW

493 40.Baek, M. et al. Accurate prediction of protein–nucleic acid complexes using RoseTTAFoldNA.

494 Nat. Methods 1–5 (2023).

495 41.Chen, K., Zhou, Y., Wang, S. & Xiong, P. RNA tertiary structure modeling with BRiQ potential

496 in CASP15. Proteins: Struct. Funct. Bioinf. 91, 1771–1778 (2023).

497 42.Basu, S. & Wallner, B. DockQ: A Quality Measure for Protein-Protein Docking Models. PLoS

498 One 11, e0161879 (2016).

499 43.Zhang, Y. & Skolnick, J. Scoring function for automated assessment of protein structure template

500 quality. Proteins 57, 702–710 (2004).

501 44.Watson, E. R. et al. Molecular glue CELMoD compounds are regulators of cereblon

502 conformation. Science 378, 549–553 (2022).

503 45.Wayment-Steele, H. K. et al. Predicting multiple conformations via sequence clustering and

504 AlphaFold2. Nature 1–3 (2023).

505 46.del Alamo, D., Sala, D., Mchaourab, H. S. & Meiler, J. Sampling alternative conformational

506 states of transporters and receptors with AlphaFold2. (2022) doi:10.7554/eLife.75751.

507 47.Heo, L. & Feig, M. Multi-state modeling of G-protein coupled receptors at experimental

508 accuracy. Proteins 90, 1873–1885 (2022).

509 48.Wallner, B. AFsample: improving multimer prediction with AlphaFold using massive sampling.

510 Bioinformatics 39, (2023).

511 Methods

512 Full algorithm details

513 Extensive explanations of the components are available in Supplementary Methods 2–5 In

514 addition, pseudocode is available in Supplementary Algorithms 1–31, network diagrams in

515 Fig. 1d, Fig. 2a,b,c, and Supplementary Fig. 2, input features in Supplementary Table 5, and

516 additional hyper parameters for training in Supplementary Tables 3, 4, 7.

ACCELERATED ARTICLE PREVIEW

517 Training regime

518 No structural data used during training was released after 2021-09-30, and for the model used in

PoseBusters evaluations we filtered out PDB32 519 structures released after 2019-09-30. One

520 optimizer step uses a mini batch of 256 input data samples and during initial training 256 * 48 =

521 12,288 diffusion samples. For fine tuning the number of diffusion samples is reduced to 256 * 32

522 = 8,192. The model is trained in three stages, the initial training with a crop size of 384 tokens

523 and two sequential fine tuning stages with crop sizes 640 and 768 tokens. See Supplementary

524 Methods 5.2 for more details.

525 Inference regime

526 No inference time templates or reference ligand position features were released after 2021-09-30,

527 and in the case of PoseBusters evaluation, an earlier cutoff date of 2019-09-30 was used. The

528 model can be run with different random seeds to generate alternative results, with a batch of

529 diffusion samples per seed. Unless otherwise stated, all results are generated by selecting the top

530 confidence sample from running 5 seeds of the same trained model, with 5 diffusion samples per

531 model seed, for a total of 25 samples to choose from. Standard crystallisation aids are excluded

532 from predictions (see Supplementary Table 8).

533 Results are shown for the top ranked sample and sample ranking depends on whether trying to

534 select the overall best output globally, or the best output for some chain, interface or modified

535 residue. Global ranking uses a mix of pTM and ipTM along with terms to reduce cases with large

536 numbers of clashes and increase rates of disorder, individual chain ranking uses a chain specific

537 pTM measure, interface ranking uses a bespoke ipTM measure for the relevant chain pair and

538 modified residue ranking uses average pLDDT over the residue of interest (see Supplementary

539 Methods 5.9.3 for details).

540 Metrics

541 Evaluation compares a predicted structure to the corresponding ground truth structure. If the

542 complex contains multiple identical entities, assignment of the predicted units to the ground truth

543 units is found by maximising LDDT. Assignment in local symmetry groups of atoms in ligands

544 is solved by exhaustive search over the first 1000 per-residue symmetries as given by RDKit.

545 We measure the quality of the predictions with DockQ, LDDT or pocket-aligned RMSD. For

546 nucleic-protein interfaces we measure interface accuracy via interface LDDT (iLDDT), which is

547 calculated from distances between atoms across different chains in the interface. DockQ and

548 iLDDT are highly correlated (Extended Data Fig. 9), so the standard cutoffs for DockQ can be

549 translated to equivalent iLDDT cutoffs. Nucleic acid LDDTs (intra-chains and interface) were

550 calculated with an inclusion radius of 30 Å compared to the usual 15 Å used for proteins, owing

551 to their larger scale. For confidence calibration assessment, we use a bespoke LDDT,

552 “LDDT_to_polymer” metric which considers differences from each atom of a given entity to any

ACCELERATED ARTICLE PREVIEW

C

α 553 or C1’ polymer atom within its inclusion radius. This is closely related to how the confidence

554 prediction is trained (see Supplementary Methods 4.3.1 for details).

555 Pocket-aligned RMSD is computed as follows: the pocket is defined as all heavy atoms within

556 10 Å of any heavy atom of the ligand, restricted to the primary polymer chain for the ligand or

557 modified residue being scored, and further restricted to only backbone atoms for proteins. The

558 primary polymer chain is defined variously: for PoseBusters it is the protein chain with the most

559 atoms within 10 Å of the ligand, for bonded ligand scores it is the bonded polymer chain and for

560 modified residues it is the chain that the residue is contained in (minus that residue). The pocket

561 is used to align the predicted structure to the ground truth structure with least squares rigid

562 alignment and then RMSD is computed on all heavy atoms of the ligand.

563 Recent PDB evaluation set

564 General model evaluation was performed on our Recent PDB set consisting of 8,856 PDB

565 complexes released between 2022-05-01 and 2023-01-12. The set contains almost all PDB

566 complexes released during that period less than 5,120 model tokens in size (see Supplementary

567 Methods 6.1 for details). Single chains and interfaces within each structure were scored

568 separately rather than only looking at full complex scores, then clustering was applied to chains

569 and interfaces so that scores could be aggregated first within clusters and then across clusters for

570 mean scores, or using a weighting of inverse cluster size for distributional statistics (see

571 Supplementary Methods 6.2 and 6.4 for details).

572

573 Evaluation on ligands excludes standard crystallisation aids (Supplementary Table 8), our

574 ligand exclusion list (Supplementary Table 9) and glycans (Supplementary Table 10). Bonded

575 and non-bonded ligands are evaluated separately. Ions are only included when specifically

576 mentioned (see Supplementary Table 11).

577

578 The Recent PDB set is filtered to a low homology subset (see Supplementary Methods 6.1) for

579 some results where stated. Homology is defined as sequence identity to sequences in the training

580 set and is measured via template search (see Supplementary Methods 2.4 for details).

581 Individual polymer chains in evaluation complexes are filtered out if the maximum sequence

582 identity to chains in the training set is greater than 40%, where sequence identity is the percent of

583 residues in the evaluation set chain that are identical to the training set chain. Individual peptide

584 chains (protein chains with less than 16 residues) are always filtered out. For polymer-polymer

585 interfaces, if both polymers have greater than 40% sequence identity to two chains in the same

586 complex in the training set, then the interface is filtered out. For interfaces to a peptide the

587 interface is filtered out if the non-peptide entity has greater than 40% sequence identity to any

588 chain in the training set.

589

590 To compare quality of prediction of protein-protein interfaces and protein monomers against that

of AlphaFold-Multimer v2.3 (AF-M 2.38 591 ), and to compare dependence of single protein chain

592 prediction quality on MSA depth, we restrict the low homology Recent PDB set to complexes

ACCELERATED ARTICLE PREVIEW

593 with fewer than 20 protein chains and fewer than 2,560 tokens. We compare against unrelaxed

594 AF-M 2.3 predictions.

595

596 To study antibody-antigen interface prediction, we filter the low homology Recent PDB set to

597 complexes that contain at least one protein-protein interface where one of the protein chains is in

598 one of the two largest PDB chain clusters (these clusters are representative of antibodies). We

599 further filter to complexes with at most 2,560 tokens and with no unknown amino acids in PDB,

600 to allow extensive comparison against relaxed predictions of AlphaFold-Multimer v2.3. That

601 leaves 71 antibody-antigen complexes, containing 166 antibody-antigen interfaces spanning 65

602 interface clusters.

603

604 MSA depth analysis (Extended Data Fig. 7a) was based on computing the normalised number

605 of effective sequences (Neff) for each position of a query sequence. Per-residue Neff values were

606 obtained by counting the number of non-gap residues in the MSA for this position and weighting

the sequences using the Neff scheme49 607 with a threshold of 80% sequence identity measured on the

608 region that is non-gap in either sequence.

609 Nucleic acid prediction baseline

610 For benchmarking performance on nucleic acid structure prediction, we report baseline

611 comparisons to an existing machine learning system for protein-nucleic acid and RNA tertiary

structure prediction, RoseTTAFold2NA18. We run the open source RF2NA50 612 with the same

613 multiple sequence alignments (MSAs) as were used for AlphaFold 3 predictions. For comparison

614 between AlphaFold 3 and RF2NA, a subset of our Recent PDB set are chosen to meet the

615 RF2NA criteria (<1000 total residues and nucleotides). As RF2NA was not trained to predict

616 systems with DNA and RNA, analysis is limited to targets with only one nucleic acid type. No

617 system was publically available at time of writing for baseline comparisons on data with

618 arbitrary combinations of biomolecular types in PDB.

619

620 As an additional baseline for RNA tertiary structure prediction, we evaluate AlphaFold 3

621 performance on CASP15 RNA targets that are currently publicly available (R1116/8S95,

622 R1117/8FZA, R1126 (downloaded from the CASP 15 website

623 https://predictioncenter.org/casp15/TARGETS_PDB/R1126.pdb), R1128/8BTZ, R1136/7ZJ4,

624 R1138/[7PTK/7PTL], R1189/7YR7, and R1190/7YR6). We compare top-1 ranked predictions,

625 and where multiple ground truth structures exist (R1136) the prediction is scored against the

626 closest state. We display comparisons to RF2NA as a representative machine learning system,

627 AIchemy_RNA2 as the top performing entrant with human intervention, and AIchemy_RNA as

628 the top performing machine learning system. All entrants’ predictions were downloaded from the

629 CASP website and scored internally.

ACCELERATED ARTICLE PREVIEW

630 PoseBusters

631 While other analyses used an AlphaFold model trained on PDB data released prior to a cutoff of

632 2021-09-30, our PoseBusters analysis was conducted on a model (with identical architecture and

633 similar training schedule) differing only in the use of an earlier 2019-09-30 cutoff. This analysis

634 therefore did not include training data, inference time templates, or “ref_pos” features released

635 after this date.

636

637 Inference was performed on the asymmetric unit from specified PDBs, with the following minor

638 modifications. In several PDB files, chains clashing with the ligand of interest were removed

639 (7O1T, 7PUV, 7SCW, 7WJB, 7ZXV, 8AIE). Another PDB (8F4J) was too large to inference the

640 entire system (over 5120 tokens), so we only included protein chains within 20 Å of the ligand of

641 interest. Five model seeds, each with five diffusion samples, were produced per target, resulting

642 in 25 predictions, which were ranked by quality and predicted accuracy: the ranking score was

643 calculated from an ipTM aggregate (Supplementary Methods 5.9.3 point 3), then further

644 divided by 100 if the ligand had chirality errors or had clashes with the protein.

645

646 For pocket-aligned RMSD, first alignment between the predicted and ground truth structures was

647 conducted by aligning to ground truth pocket backbone atoms (CA, C, or N atoms within 10 Å of

648 the ligand of interest) from the primary protein chain (the chain with the greatest number of

contacts within 10 Å of the ligand). The posebusters python package v0.2.751 649 was used to score

650 RMSD and violations from the pocket-aligned predictions.

651

652 While AlphaFold models are “blind” to the protein pocket, docking is often performed with

653 knowledge of the protein pocket residues. For example, Uni-Mol specifies the pocket as any

residue within 6 Å of the heavy atoms in the ligand of interest26 654 . To evaluate the ability of

655 AlphaFold 3 to “dock” ligands accurately when given pocket information, we fine-tuned a 2019-

656 09-30-cutoff AlphaFold 3 model with an additional token feature specifying pocket-ligand pairs

657 (Supplementary Methods 2.8). Specifically, an additional token feature was introduced, set to

658 true for a ligand entity of interest and any pocket residues with heavy atoms within 6 Å of the

659 ligand entity. At training time a single random ligand entity is chosen to use in this feature. Note

660 that multiple ligand chains with the same entity (CCD code) may be selected. At inference time,

661 the ligand entity was chosen based on the ligand of interest’s CCD code, so again multiple ligand

662 chains were occasionally chosen. Results of this analysis are shown in Extended Data Fig. 4.

663 Model Performance Analysis and Visualization

664 Data analysis used Python v3.11.7 (https://www.python.org/), NumPy v1.26.3

665 (https://github.com/numpy/numpy), SciPy v1.9.3 (https:// www.scipy.org/), seaborn v0.12.2

666 (https://github.com/mwaskom/seaborn), Matplotlib v3.6.1

667 (https://github.com/matplotlib/matplotlib), pandas v2.0.3 (https://github.com/pandas668 dev/pandas), statsmodels v0.12.2 (https://github.com/statsmodels/statsmodels), RDKit v4.3.0

669 (https://github.com/rdkit/rdkit), and Colab (https://research.google.com/colaboratory). TM-align

ACCELERATED ARTICLE PREVIEW

670 v20190822 (https://zhanglab.dcmb.med.umich.edu/TM-align/) was used for computing TM671 scores. Structure visualizations were created in Pymol v2.55.5

672 (https://github.com/schrodinger/pymol-open-source).

673

674

675 Data availability

676 All scientific datasets used to create training and evaluation inputs are freely available from

677 public sources. Structures from the PDB were used for training and as templates

678 (https://files.wwpdb.org/pub/pdb/data/assemblies/mmCIF/; for sequence clusters see

679 https://cdn.rcsb.org/resources/sequence/clusters/clusters-by-entity-40.txt; for sequence data see

680 https://files.wwpdb.org/pub/pdb/derived_data/).

681 Training used a version of the PDB downloaded 12 January 2023, while template search used a

682 version downloaded 28 September 2022. We also used the Chemical Components Dictionary

683 downloaded on 19 October 2023 (https://www.wwpdb.org/data/ccd).

We show experimental structures from the PDB with accession numbers 7PZB52,53, 7PNM54,55 684 ,

7TQL56,57, 7AU258,59, 7U8C60,61, 7URD62,63, 7WUX64,65, 7QIE66,67, 7T8268,69, 7CTM70,71 685 ,

8CVP44,72, 8D7U44,73, 7F6074,75, 8BTI76,77, 7KZ978,79, 7XFA80,81, 7PEU82,83, 7SDW84,85 686 ,

7TNZ86,87, 7R6R 88,89, 7USR90,91, and 7Z1K.92,93 687

688

689 We also used the following publicly available databases for training or evaluation. Detailed

690 usage is described in Supplementary Methods 2.2 and Supplementary Methods 2.5.2.

691 UniRef90 v.2020_01 (https://ftp.ebi.ac.uk/pub/databases/uniprot/previous_releases/release692 2020_01/uniref/),

693 UniRef90 v.2020_03 (https://ftp.ebi.ac.uk/pub/databases/uniprot/previous_releases/release694 2020_03/uniref/),

695 UniRef90 v.2022_05

696 https://ftp.ebi.ac.uk/pub/databases/uniprot/previous_releases/release-2022_05/uniref/),

697 Uniclust30 v.2018_08

698 (https://wwwuser.gwdg.de/~compbiol/uniclust/2018_08/),

699 Uniclust30 v.2021_03

700 (https://wwwuser.gwdg.de/~compbiol/uniclust/2021_03/),

701 MGnify clusters v.2018_12

702 (https://ftp.ebi.ac.uk/pub/databases/metagenomics/peptide_database/2018_12/),

703 MGnify clusters v.2022_05

704 (https://ftp.ebi.ac.uk/pub/databases/metagenomics/peptide_database/2022_05/),

705 BFD

706 (https://bfd.mmseqs.com),

707 RFam v.14.9

708 (https://ftp.ebi.ac.uk/pub/databases/Rfam/14.9/),

709 RNAcentral v.21.0

ACCELERATED ARTICLE PREVIEW

710 (https://ftp.ebi.ac.uk/pub/databases/RNAcentral/releases/21.0/),

711 Nucleotide Database (as of 23 February 2023)

712 (https://ftp.ncbi.nlm.nih.gov/blast/db/FASTA/nt.gz),

713 JASPAR 2022

714 (https://jaspar.elixir.no/downloads/; see https://jaspar.elixir.no/profile-versions for version

715 information),

SELEX protein sequences from Supplementary Tables94 716

717 (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8009048/),

SELEX protein sequences from Supplementary Tables95 718

719 (https://www.nature.com/articles/nature15518).

720

721 Code availability

722 AlphaFold 3 will be available as a non-commercial usage only server at

723 https://www.alphafoldserver.com, with restrictions on allowed ligands and covalent

724 modifications. Pseudocode describing the algorithms is available in the Supplementary

725 Information. Code is not provided.

726 Acknowledgements

727 We thank Giorgio Arena, Žiga Avsec, Anthony Baryshnikov, Russ Bates, Molly Beck, Aliyah

728 Bond, Nathalie Bradley-Schmieg, Jana Cavojska, Ben Coppin, Emilien Dupont, Sean Eddy,

729 Marco Fiscato, Richard Green, Dhavanthi Hariharan, Kristian Holsheimer, Nicole Hurley, Chris

730 Jones, Koray Kavukcuoglu, Jacob Kelly, Eugenia Kim, Anna Koivuniemi, Oleg Kovalevskiy,

731 Dariusz Lasecki, Meera Last, Agata Laydon, William McCorkindale, Sam Miller, Alex Morris,

732 Lauren Nicolaisen, Evan Palmer, Antonia Paterson, Stig Petersen, Ollie Purkiss, Chongyang Shi,

733 George Thomas, Gregory Thornton and Hamish Tomlinson for their contributions

734 Author Contributions

735 The equally contributing authors are alphabetically ordered, as are the remaining core contributor

736 authors (excluding jointly supervising authors) and similar for all remaining non-supervising

737 authors. D.H., M.J. and J.J. led the research. M.J., J.J. and P.K. developed research strategy. J.

738 Abramson, V.B., T.G. and C.-C.H. led key research pillars. T.G. and A. Žídek led the technical

739 framework for research. O.B., H.G. and S.S. coordinated and managed the Research Project. J.

740 Abramson, J. Adler, E.A., A. Ballard, J.B., V.B., A.C.-R., J.D., R.E., D.A.E., M.F., F.F., T.G.,

741 C.-C.H., M.J., J.J., Y.A.K., A. Potapenko, A. Pritzel, D.R., O.R., A.T., C.T., K.T., L.W., Z.W.

742 and E.Z. developed the neural network architecture and training procedure. J. Abramson, A.

743 Ballard, J.B., V.B., C.B., S.B., A. Bridgland, A. Cherepanov, A.C.-R., A. Cowie, J.D., T.G., R.J.,

744 M.O., K.P., D.R., O.R., M.Z., A. Žemgulytė and A. Žídek developed the training, inference,

745 data, and evaluation infrastructure. J. Abramson, J. Adler, A. Ballard, V.B., A.C.-R., R.E.,

ACCELERATED ARTICLE PREVIEW

746 D.A.E., T.G., D.H., M.J., J.J., P.K., K.P., A. Pritzel, O.R., P.S., S.S., A.S., K.T. and L.W.

747 contributed to the writing of the paper. M.C., C.M.R.L., S.Y. advised on the project.

748 Competing Interests

749 Author-affiliated entities have filed US provisional patent applications including 63/611,674,

750 63/611,638 and 63/546,444 relating to predicting three-dimensional (3d) structures of molecule

751 complexes using embedding neural networks and generative models. All authors other than A.

752 Bridgland, Y.A.K. and E.Z. have commercial interests in the work described.

753 Further Citations

754 49.Wu, T., Hou, J., Adhikari, B. & Cheng, J. Analysis of several key factors influencing deep

755 learning-based inter-residue contact prediction. Bioinformatics 36, 1091–1098 (2020).

756 50.Release RF2NA v0.2 · uw-ipd/RoseTTAFold2NA. GitHub https://github.com/uw757 ipd/RoseTTAFold2NA/releases/tag/v0.2.

758 51.Release v0.2.7 · maabuu/posebusters. GitHub

759 https://github.com/maabuu/posebusters/releases/tag/v0.2.7.

760 52.Werel, L. et al. Structural Basis of Dual Specificity of Sinorhizobium meliloti Clr, a cAMP and

761 cGMP Receptor Protein. MBio 14, e0302822 (2023).

762 53.wwPDB: 7PZB. https://doi.org/10.2210/pdb7PZB/pdb.

763 54.Wang, C. et al. Antigenic structure of the human coronavirus OC43 spike reveals exposed and

764 occluded neutralizing epitopes. Nat. Commun. 13, 1–15 (2022).

765 55.wwPDB: 7PNM. https://doi.org/10.2210/pdb7PNM/pdb.

766 56.Lapointe, C. P. et al. eIF5B and eIF1A reorient initiator tRNA to allow ribosomal subunit joining.

767 Nature 607, 185–190 (2022).

768 57.wwPDB: 7TQL. https://doi.org/10.2210/pdb7TQL/pdb.

769 58.Wilson, L. F. L. et al. The structure of EXTL3 helps to explain the different roles of bi-domain

770 exostosins in heparan sulfate synthesis. Nat. Commun. 13, 1–15 (2022).

771 59.wwPDB: 7AU2. https://doi.org/10.2210/pdb7AU2/pdb.

772 60.Liu, X. et al. Highly active CAR T cells that bind to a juxtamembrane region of mesothelin and

ACCELERATED ARTICLE PREVIEW

773 are not blocked by shed mesothelin. Proc. Natl. Acad. Sci. U. S. A. 119, e2202439119 (2022).

774 61.wwPDB: 7U8C. https://doi.org/10.2210/pdb7U8C/pdb.

775 62.Liu, Y. et al. Mechanisms and inhibition of Porcupine-mediated Wnt acylation. Nature 607, 816–

776 822 (2022).

777 63.wwPDB: 7URD. https://doi.org/10.2210/pdb7URD/pdb.

778 64.Kurosawa, S. et al. Molecular Basis for Enzymatic Aziridine Formation via Sulfate Elimination.

779 J. Am. Chem. Soc. 144, 16164–16170 (2022).

780 65.wwPDB: 7WUX. https://doi.org/10.2210/pdb7WUX/pdb.

781 66.Boffey, H. K. et al. Development of Selective Phosphatidylinositol 5-Phosphate 4-Kinase γ

782 Inhibitors with a Non-ATP-competitive, Allosteric Binding Mode. J. Med. Chem. 65, 3359–3370

783 (2022).

784 67.wwPDB: 7QIE. https://doi.org/10.2210/pdb7QIE/pdb.

785 68.Buckley, P. T. et al. Multivalent human antibody-centyrin fusion protein to prevent and treat

786 Staphylococcus aureus infections. Cell Host Microbe 31, 751–765.e11 (2023).

787 69.wwPDB: 7T82. https://doi.org/10.2210/pdb7T82/pdb.

788 70.Mohapatra, S. B. & Manoj, N. Structural basis of catalysis and substrate recognition by the

789 NAD(H)-dependent α-d-glucuronidase from the glycoside hydrolase family 4. Biochem. J 478, 943–

790 959 (2021).

791 71.wwPDB: 7CTM. https://doi.org/10.2210/pdb7CTM/pdb.

792 72.wwPDB: 8CVP. https://doi.org/10.2210/pdb8CVP/pdb.

793 73.wwPDB: 8D7U. https://doi.org/10.2210/pdb8D7U/pdb.

794 74.Gao, X. et al. Structural basis for Sarbecovirus ORF6 mediated blockage of nucleocytoplasmic

795 transport. Nat. Commun. 13, 1–11 (2022).

796 75.wwPDB: 7F60. https://doi.org/10.2210/pdb7F60/pdb.

797 76.Atkinson, B. N. et al. Designed switch from covalent to non-covalent inhibitors of

798 carboxylesterase Notum activity. Eur. J. Med. Chem. 251, 115132 (2023).

799 77.wwPDB: 8BTI. https://doi.org/10.2210/pdb8BTI/pdb.

ACCELERATED ARTICLE PREVIEW

800 78.Luo, S. et al. Structural basis for a bacterial Pip system plant effector recognition protein. Proc.

801 Natl. Acad. Sci. U. S. A. 118, (2021).

802 79.wwPDB: 7KZ9. https://doi.org/10.2210/pdb7KZ9/pdb.

803 80.Liu, C. et al. Identification of Monosaccharide Derivatives as Potent, Selective, and Orally

804 Bioavailable Inhibitors of Human and Mouse Galectin-3. J. Med. Chem. 65, 11084–11099 (2022).

805 81.wwPDB: 7XFA. https://doi.org/10.2210/pdb7XFA/pdb.

806 82.Dombrowski, M., Engeholm, M., Dienemann, C., Dodonova, S. & Cramer, P. Histone H1

807 binding to nucleosome arrays depends on linker DNA length and trajectory. Nat. Struct. Mol. Biol.

808 29, 493–501 (2022).

809 83.wwPDB: 7PEU. https://doi.org/10.2210/pdb7PEU/pdb.

810 84.Vecchioni, S. et al. Metal-Mediated DNA Nanotechnology in 3D: Structural Library by

811 Templated Diffraction. Adv. Mater. 35, e2210938 (2023).

812 85.wwPDB: 7SDW. https://doi.org/10.2210/pdb7SDW/pdb.

813 86.Wang, W. & Pyle, A. M. The RIG-I receptor adopts two different conformations for

814 distinguishing host from viral RNA ligands. Mol. Cell 82, 4131–4144.e6 (2022).

815 87.wwPDB: 7TNZ. https://doi.org/10.2210/pdb7TNZ/pdb.

816 88.McGinnis, R. J. et al. A monomeric mycobacteriophage immunity repressor utilizes two domains

817 to recognize an asymmetric DNA sequence. Nat. Commun. 13, 4105 (2022).

818 89.wwPDB: 7R6R. https://doi.org/10.2210/pdb7R6R/pdb.

819 90.Dietrich, M. H. et al. Nanobodies against Pfs230 block Plasmodium falciparum transmission.

820 Biochem. J 479, 2529–2546 (2022).

821 91.wwPDB: 7USR. https://doi.org/10.2210/pdb7USR/pdb.

822 92.Appel, L.-M. et al. The SPOC domain is a phosphoserine binding module that bridges

823 transcription machinery with co- and post-transcriptional regulators. Nat. Commun. 14, 1–22 (2023).

824 93.wwPDB: 7Z1K. https://doi.org/10.2210/pdb7Z1K/pdb.

825 94.Yin, Y. et al. Impact of cytosine methylation on DNA binding specificities of human

826 transcription factors. Science 356, (2017).

ACCELERATED ARTICLE PREVIEW

827 95.Jolma, A. et al. DNA-dependent formation of transcription factor pairs alters their binding

828 specificity. Nature 527, 384–388 (2015).

829 Extended Data Figure Captions

830

831 Extended Data Figure 1 | Disordered region prediction. a, Example prediction for a disordered protein from

832 AlphaFoldMultimer v2.3, AlphaFold 3, and AlphaFold 3 trained without the disordered protein PDB cross

833 distillation set. Protein is DP02376 from the CAID 2 (Critical Assessment of protein Intrinsic Disorder prediction)

834 set. Predictions coloured by pLDDT (orange: pLDDT<=50 ,yellow: 50<pLDDT<=70, light blue: 70<pLDDT<=90,

835 and dark blue: 90<=pLDDT<100). b, Predictions of disorder across residues in proteins in the CAID 2 set, which

836 are also low homology to the AF3 training set. Prediction methods include RASA (relative accessible surface area)

837 and pLDDT (N=151 proteins; 46,093 residues).

838

839 Extended Data Figure 2 | Accuracy across training. Training curves for initial training and fine tuning showing

840 LDDT (local distance difference test) on our evaluation set as a function of optimizer steps. One optimizer step uses

841 a mini batch of 256 trunk samples and during initial training 256 * 48 = 12,288 diffusion samples. For fine tuning

842 the number of diffusion samples is reduced to 256 * 32 = 8,192. The scatter plot shows the raw data points and the

843 lines show the smoothed performance using a median filter with a kernel width of 9 data points. The dashed lines

844 mark the points where the smoothed performance passes 90% and 97% of the initial training maximum for the first

845 time.

846

847 Extended Data Figure 3 | AlphaFold 3 predictions of PoseBusters examples for which Vina and Gold were

848 inaccurate. Predicted protein chains are shown in blue, predicted ligands in orange, and ground truth in grey. a,

849 Human Notum bound to inhibitor ARUK3004556 (PDB ID 8BTI, ligand RMSD: 0.65 Å). b, Pseudomonas sp.

850 PDC86 Aapf bound to HEHEAA (PDB ID 7KZ9, ligand RMSD: 1.3 Å). c, Human Galectin-3 carbohydrate851 recognition domain in complex with compound 22 (PDB ID 7XFA, ligand RMSD: 0.44 Å).

852

853 Extended Data Figure 4 | PoseBusters analysis. a, Comparison of AlphaFold 3 and baseline method protein854 ligand binding success on the PoseBusters Version 1 benchmark set (V1, August 2023 release). Methods classified

855 by the extent of ground truth information used to make predictions. Note all methods that use pocket residue

856 information except for UMol and AF3 also use ground truth holo protein structures. b, PoseBusters Version 2 (V2,

857 November 2023 release) comparison between the leading docking method Vina and AF3 2019 (two-sided Fisher

exact test, N = 308 targets, p = 2.3 * 10−8 858 ). c, PoseBusters V2 results of AF3 2019 on targets with low, moderate,

859 and high protein sequence homology (integer ranges indicate maximum sequence identity with proteins in the

860 training set). d, PoseBusters V2 results of AF3 2019 with ligands split by those characterised as “common natural”

861 ligands and others. “Common natural” ligands are defined as those which occur greater than 100 times in the PDB

862 and which are not non-natural (by visual inspection). A full list may be found in Supplementary Table 15. Dark bar

863 indicates RMSD < 2 Å and passing PoseBusters validity checks (PB-valid). e, PoseBusters V2 structural accuracy

864 and validity. Dark bar indicates RMSD < 2 Å and passing PoseBusters validity checks (PB-valid). Light hashed bar

865 indicates RMSD < 2 Å but not PB valid. f, PoseBusters V2 detailed validity check comparison. Error bars indicate

866 exact binomial distribution 95% confidence intervals. N=427 targets for RoseTTAFold All-Atom and 428 targets for

867 all others in Version 1; 308 targets in Version 2.

868

869 Extended Data Figure 5 | Nucleic acid prediction accuracy and confidences. a, CASP15 RNA prediction

870 accuracy from AIChemy_RNA (the top AI-based submission), RoseTTAFold2NA (the AI-based method capable of

871 predicting proteinRNA complexes), and AlphaFold 3. Ten of the 13 targets are available in the PDB or via the

872 CASP15 website for evaluation. Predictions are downloaded from the CASP website for external models. b,

ACCELERATED ARTICLE PREVIEW

873 Accuracy on structures containing low homology RNA-only or DNA-only complexes from the recent PDB

874 evaluation set. Comparison between AlphaFold 3 and RoseTTAFold2NA (RF2NA) (RNA: N=29 structures, paired

Wilcoxon signed-rank test, p=1.6 * 10−7 875 ; DNA: N=63 structures, paired two-sided Wilcoxon signed-rank test, p=5.2

* 10−12 876 ). Note RF2NA was only trained and evaluated on duplexes (chains forming at least 10 hydrogen bonds), but

877 some DNA structures in this set may not be duplexes. Box, centerline, and whiskers boundaries are at (25%, 75%)

878 intervals, median, and (5%, 95%) intervals. c Predicted structure of a mycobacteriophage immunity repressor

879 protein bound to double stranded DNA (PDB ID 7R6R), coloured by pLDDT (left; orange: 0-50, yellow: 50-70,

880 cyan 70-90, and blue 90-100) and chain id (right). Note the disordered N-terminus not entirely shown. d, Predicted

881 aligned error (PAE) per token-pair for the prediction in c with rows and columns labelled by chain id and green

882 gradient indicating PAE.

883

884 Extended Data Figure 6 | Analysis and examples for modified proteins and nucleic acids. a, Accuracy on

885 structures

886 containing common phosphorylation residues (SEP, TPO, PTR, NEP, HIP) from the recent PDB evaluation set.

887 Comparison between AlphaFold 3 with phosphorylation modelled, and AlphaFold 3 without modelling

phosphorylation (N=76 clusters, paired two-sided Wilcoxon signed-rank test, p=1.6 * 10−4 888 ). Note, to predict a

889 structure without modelling phosphorylation, we predict the parent (standard) residue in place of the modification.

890 AlphaFold 3 generally achieves better backbone accuracy when modelling phosphorylation. Error bars indicate

891 exact binomial distribution 95% confidence intervals. b, SPOC domain of human SHARP in complex with

892 phosphorylated RNA polymerase II C-terminal domain (PDB ID 7Z1K), predictions coloured by pLDDT (orange:

893 0-50, yellow: 50-70, cyan 70-90, and blue 90-100). Left: Phosphorylation modelled (mean pocket-aligned RMSDCα

894 2.104 Å). Right: Without modelling phosphorylation (mean pocketaligned RMSDCα 10.261 Å). When excluding

895 phosphorylation, AlphaFold 3 provides lower pLDDT confidence on the phosphopeptide. c, Structure of parkin

896 bound to two phospho-ubiquitin molecules (PDB ID 7US1), predictions similarly coloured by pLDDT. Left:

897 Phosphorylation modelled (mean pocket-aligned RMSDCα 0.424 Å). Right: Without modelling phosphorylation

898 (mean pocket-aligned RMSDCα 9.706 Å). When excluding phosphorylation, AlphaFold 3 provides lower pLDDT

899 confidence on the interface residues of the incorrectly predicted ubiquitin. d, Example structures with modified

900 nucleic acids. Left: Guanosine monophosphate in RNA (PDB ID 7TNZ, mean pocket-aligned modified residue

901 RMSD 0.840 Å). Right: Methylated DNA cytosines (PDB ID 7SDW, mean pocket-aligned modified residue RMSD

902 0.502 Å). Welabel residues of the predicted structure for reference. Ground truth structure in grey; predicted protein

903 in blue, predicted RNA in purple, predicted DNA in magenta, predicted ions in orange, with predicted modifications

904 highlighted via spheres

905

906 Extended Data Figure 7 | Model accuracy with MSA size and number of seeds. a, Effect of MSA depth on

907 protein prediction accuracy. Accuracy is given as single chain LDDT score and MSA depth is computed by counting

908 the number of non-gap residues for each position in the MSA using the Neff weighting scheme and taking the median

909 across residues (see Methods for details on Neff). MSA used for AF-M 2.3 differs slightly from AF3; the data uses

910 the AF3 MSA depth for both to make the comparison clearer. The analysis uses every protein chain in the low

911 homology Recent PDB set, restricted to chains in complexes with fewer than 20 protein chains and fewer than 2,560

912 tokens (see Methods for details on Recent PDB set and comparisons to AF-M 2.3). The curves are obtained through

913 Gaussian kernel average smoothing (window size is 0.2 units in log10(Neff)); the shaded area is the 95% confidence

914 interval estimated using bootstrap of 10,000 samples. b, Increase in ranked accuracy with number of seeds for

915 different molecule types. Predictions are ranked by confidence, and only the most confident per interface is scored.

916 Evaluated on the low homology recent PDB set, filtered to less than 1,536 tokens. Number of clusters evaluated:

917 dna-intra=386, protein-intra=875, rnaintra=78, protein-dna=307, protein-rna=102, protein-protein

918 (antibody=False)=697, protein-protein (antibody=True)=58. Confidence intervals are 95% bootstraps over 1,000

919 samples.

920

921 Extended Data Figure 8 | Relationship between confidence and accuracy for protein interactions with ions,

922 bonded ligands and bonded glycans. Accuracy is given as the percentage of interface clusters under various

923 pocket-aligned RMSD thresholds, as a function of the chain pair ipTM of the interface. The ions group includes both

ACCELERATED ARTICLE PREVIEW

924 metals and nonmetals. N values report the number of clusters in each band. For a similar analysis on general ligand925 protein interfaces, see Figure 4 of main text.

926

927 Extended Data Figure 9 |Correlation of DockQ and iLDDT for protein-protein interfaces. One data point per

928 cluster, 4,182 clusters shown. Line of best fit with a Huber regressor with epsilon 1. DockQ categories correct

929 (>0.23), and very high accuracy (>0.8) correspond to iLDDTs of 23.6 and 77.6 respectively

930

931 Extended Data Table 1 | Prediction accuracy across biomolecular complexes. AlphaFold 3 Performance on

932 PoseBusters V1 (August 2023 release), PoseBusters V2 (November 6th 2023 release), and our Recent PDB

933 evaluation set. For ligands and nucleic acids N indicates number of structures; for covalent modifications and

934 proteins N indicates number of clusters.

ACCELERATED ARTICLE PREVIEW

a b

c

AlphaFold 3

2019 cutoff

N=428

AutoDock

Vina

N=428

RoseTTAFold

All-Atom

N=427

0

20

40

60

80

100

Success (%)

***

**

Ligands PoseBusters Set

PDB

Protein-RNA

N = 25

PDB

Protein-dsDNA

N = 38

CASP 15

RNA

N = 8

**

***

Nucleic Acids

AlphaFold 3

RoseTTAFold2NA

AIchemy_RNA2 (has human input)

Sequences,

ligands,

covalent

bonds

Input

embedder

(3 blocks) Conformer

generation

Template

search

+

Template

module

(2 blocks)

MSA

module

(4 blocks) Recycling Genetic

search

+

pair

inputs

single

d

ACCELERATED AR

Bonded

ligands

N=66

Glycosylation

N=28

Protein

N=40

Modified Residues

DNA

N=91

RNA

N=23

Covalent Modifications

All

Protein-Protein

N=1064

ProteinAntibody

N=65

Protein

Monomers

N=338

***

***

***

Proteins

AlphaFold 3

AlphaFold-Multimer 2.3

Pairformer

(48 blocks) Diffusion

module

(3 + 24 + 3 blocks)

+

Diffusion iterations

Confidence

module

(4 blocks)

0 100

ARTICLE PREVIEW

pair

representation

(n,n,c)

triangle

update

using

\"outgoing\"

edges

triangle

selfattention

around

starting

node

triangle

selfattention

around

ending

node

triangle

update

using

\"incoming\"

edges

+ + + + single repr. (n,c)

48 blocks

a

global attention

(tokens)

24 blocks

seq. local

attention

(atoms)

3 blocks

seq. local

attention

(atoms)

3 blocks

peratom

cond.

pertoken

cond.

rand.

rot.

trans.

b

+

Diffusion

Module

(inference)

Network

trunk

+

20 iterations

mini rollout

48 samples

Ground

truth Permute

ground

truth

+

STOP

STSTOP

c

ACCELERATED AR

transition

pair

representation

(n,n,c)

+

single

attention

with pair

bias

single repr. (n,c)

transition

+ +

+

Confidence

Module

Metrics

Diffusion

Module

(training)

Loss

Loss

TOP

0k 20k 40k 60k 80k 100k 120k 140k

steps

30

40

50

60

70

80

90

100

LDDT

initial training

fine

tune 1

intra ligand

intra protein

intra dna

intra rna

protein-ligand

protein-protein

protein-dna

protein-rna

fine

tune 2

d

ARTICLE PREVIEW

a b d e ACCELERATED AR

c f ARTICLE PREVIEW

0-0.4

N=834

0.4-0.6

N=692

0.6-0.8

N=1574

0.8-0.95

N=1804

0.95+

N=229

chain pair ipTM

0.0

0.2

0.4

0.6

0.8

1.0

DockQ

Protein-Protein

0-0.4

N=277

0.4-0.6

N=320

0.6-0.8

N=449

0.8-0.95

N=344

0.9N=chain pair ipTM

0

20

40

60

80

100

iLDDT

Nucleic Acid-Protein

0-50

N=107

50-70

N=469

70-90

N=2283

90+

N=1552

chain pLDDT

0

20

40

60

80

100

LDDT_to_polymer

Protein

0-50

N=227

50-70

N=225

70-90

N=189

90N=2chain pLDDT

0

20

40

60

80

100

LDDT_to_polymer

Nucleic Acid

a b c

A

A

0.003

0.003

0.003

0.003

0.721

0.72C

C

0.740

0.740

D

D

F

FInterface DockQ scored

ACCELERATED AR

95+

=108

+

229

0-50

N=57

50-70

N=136

70-90

N=396

90+

N=934

chain pLDDT

0

20

40

60

80

100

LDDT_to_polymer

Ligand

0-0.4

N=7

0.4-0.6

N=29

0.6-0.8

N=92

0.8-0.95

N=299

0.95+

N=481

chain pair ipTM

0

20

40

60

80

100 % pocket RMSD < threshold

Ligand-Protein

21

F

e

0 200 400 600 800

Residue

0

200

400

600

800

Residue

A

A

C

C

D

D

F

F

Predicted Aligned Error Matrix

0

5

10

15

20

25

e 30

ARTICLE PREVIEW

1 10 100 1000

seeds per target

20

40

60 % DockQ > 0.23

***

Correct

low homology antibodies

1 10 100 1000

seeds per target

0

10

20

30

40

% DockQ > 0.8

***

Very high accuracy

low homology antibodies

AF3

AF-M 2.3

a

b

c

d Ground truth AF-M 2.3 AF3 ACCELERATED AR

*

e

ARTICLE PREVIEW

E %*1 Ѳѳ %*ѳ %*ѳ RS GVSWWHMWXMPPEXMSR

  

  

ffi!&fi$#& ' )fi%ffi'fi













%(fi$#& ' )fi%ffi'fi

 fl

 $fl

 fl

 $fl

 \"#ffl%#&&ff &' !!fl

 \"#ffl%#&&ff &' !!$fl

F

Extended Data Fig. 1

ACCELERATED ARTICLE PREVIEW

0k 20k 40k 60k 80k 100k 120k 140k

75

80

85

90

95

Intra-chain LDDT

90%

97%

Ligand

0k 20k 40k 60k 80k 100k 120k 140k

65.0

67.5

70.0

72.5

75.0

77.5

80.0

82.5

90%

97%

DNA

0k 20k 40k 60k 80k 100k 120k 140k

58

60

62

64

66

68

70

72

74

90%

97%

RNA

0k 20k 40k 60k 80k 100k 120k 140k

67.5

70.0

72.5

75.0

77.5

80.0

82.5

85.0

87.5

90%

97%

Protein

initial training

fine tuning 1 (crop: 640)

fine tuning 2 (crop: 768)

0k 20k 40k 60k 80k 100k 120k 140k

steps

50

52

54

56

58

60

62

64

Interface to protein LDDT

90%

97%

0k 20k 40k 60k 80k 100k 120k 140k

steps

44

46

48

50

52

54

56

90%

97%

0k 20k 40k 60k 80k 100k 120k 140k

steps

34

36

38

40

42

44

90%

97%

0k 20k 40k 60k 80k 100k 120k 140k

steps

48

50

52

54

56

58

60

90%

97%

Extended Data Fig. 2

ACCELERATED ARTICLE PREVIEW

E F G

Extended Data Fig. 3

ACCELERATED ARTICLE PREVIEW

RoseTTAFold

All-Atom

AF3 2019

EquiBind

TankBind

DiffDock

Vina on AF-M 2.3

DeepDock

Uni-Mol

UMol

Gold

Vina

Uni-Mol Docking V2

AF3 2019

pocket specified

0

20

40

60

80

100

% RMSD < 2Å (N=428 points)

No ground

truth struc

With holo protein

structure specified With pocket residues specified

E 4SWI&YWXIVW :IVWMSR Ϲ

Vina AF3 2019

0

20

40

60

80

100

% RMSD < 2Å (N=308 points)

***

F 4SWI&YWXIVW :IVWMSR Ѳ

(0, 30]

N=38

(30, 95]

N=83

(95, 100]

N=187

Homology cutoffs

0

20

40

60

80

100

% RMSD < 2Å

AF3 2019

G 4SWI&YWXIVW :IVWMSR Ѳ

Common

natural

ligands

N=50

Others

N=258

0

20

40

60

80

100

% RMSD < 2Å

92.0

82.0 78.3

71.3

H 4SWI&YWXIVW :IVWMSR Ѳ

AF3 2019

EquiBind

TankBind

DiffDock

DeepDock

Uni-Mol

Gold

Vina

AF3 2019

pocket specified

0

20

40

60

80

100

% RMSD < 2Å

80.5

73.1

1.9

15.9

3.2

38.0

12.7 19.5

5.2

21.8

1.9

58.1

54.5

59.7

58.1

93.2

84.4

No ground

truth struc

With holo protein

structure specified With pocket residues specified

RMSD<2

RMSD<2 and PB-valid

I 4SWI&YWXIVW :IVWMSR Ѳ

File loads

Sanitization

Molecular formula

Bonds

Tetrahedral chirality

Double bond stereochemistry

Bond lengths

Bond angles

Planar aromatic rings

Planar double bonds

Internal steric clash

Energy ratio

Minimum protein-ligand distance

Minimum distance to organic cofactors

Minimum distance to inorganic cofactors

Volume overlap with protein

Volume overlap with organic cofactors

Volume overlap with inorganic cofactors

0

20

40

60

80

100 % of all outputs passing check

Chemical validity &

consistency

Intramolecular

validity

Intermolecular

validity

DiffDock

AF3 2019 cutoff

J 4SWI&YWXIVW :IVWMSR Ѳ

Extended Data Fig. 4

ACCELERATED ARTICLE PREVIEW

E















$&fi')\"(' &$#

ffl!fi#*fl

$&fi')\"(' &$#

$'fi$\"ff

 

\"%!ffi$\"ff

ffl!fi#*fl

$'fi$\"ff













'ffl$&fi



 

 

 



  















F



\"! )

%&$'fl&'$ffl%





\"! )

%&$'fl&'$ffl%

















 

&$'fl&'$ffl%(fi&ff'fl fflfiflflfiffi%! )

 #ff\" ffi*

\"%ffl\" ffi

G H

Extended Data Fig. 5

ACCELERATED ARTICLE PREVIEW

ffl&ffi%\"flfififl



ffl&ffi\"'&%\"flfififl















*\"fffl&

)



E F \" \"!ffi\"%#ffi\"$(fi&flfl%ffl'fl%

G

H

Extended Data Fig. 6

ACCELERATED ARTICLE PREVIEW

Extended Data Fig. 7

ACCELERATED ARTICLE PREVIEW

Extended Data Fig. 8

ACCELERATED ARTICLE PREVIEW

Extended Data Fig. 9

ACCELERATED ARTICLE PREVIEW

8EWO (EXEWIX 1IXVMG 2SXIW 1IXLSH 2 1IER ѵѴ '-

0MKERHW 4SWI&YWXIVW :Ϲ 617( Ѳ ˆ ̸ 6SWI88%*SPH %PP%XSQ ѴѲ΃ ѴѲѷ ѳ΃Ѳ ̸ Ѵffi

%*ѳ ѲѷϹѶ GYXSJJ

ѴѲffi ΃ѳ ΃ѲϹ ̸ ffiѷѳ

,SPS TVSXIMR WXVYGX KMZIR )UYM&MRH ѴѲffi Ѳ Ϲѳ ̸ Ѵ

8ERO&MRH ѴѲffi Ϲѵѷ ϹϹ΃ ̸ Ϲffi΃

(MJJ(SGO ѴѲffi ѳ΃Ѷ ѳѳѲ ̸ ѴѲ

4SGOIX VIWMHYIW WTIGM͸IH :MRE SR %*1 Ѳѳ ѴѲffi ϹѳϹ Ϲѷѷ ̸ Ϲ΃

(IIT(SGO ѴѲffi Ϲ΃ffi ϹѴѳ ̸ ѲϹ΃

9RM1SP ѴѲffi ѲѲѶ ϹѶѷ ̸ Ѳ΃Ѳ

91SP ѴѲffi Ѵѵѷ Ѵѷѳ ̸ ѴѶѶ

+SPH ѴѲffi ѵϹѲ Ѵѳ ̸ ѵѷ

:MRE ѴѲffi ѵѲѳ Ѵ΃ѵ ̸ ѵ΃Ѳ

9RM1SP (SGOMRK :Ѳ ѴѲffi ΃΃ ΃ѳѳ ̸ ffiϹѴ

%*ѳ ѲѷϹѶ GYXSJJ

TSGOIX WTIGM͸IH ѴѲffi ѵѶѱ ffi΃ѷ ̸ ѶѲffi

4SWI&YWXIVW :Ѳ 617( Ѳ ˆ ̸ %*ѳ ѲѷϹѶ GYXSJJ

ѳѷffi ffiѶѴ ΃ѵ ̸ ffiѴffi

,SPS TVSXIMR WXVYGX KMZIR )UYM&MRH ѳѷffi ϹѶ ѷ΃ ̸ ѴѲ

8ERO&MRH ѳѷffi ϹѵѶ ϹѲѷ ̸ Ѳѷѵ

(MJJ(SGO ѳѷffi ѳffiѷ ѳѲѵ ̸ Ѵѳ΃

4SGOIX VIWMHYIW WTIGM͸IH :MRE SR %*1 Ѳѳ ѳѷffi Ϲѵѳ ϹϹѴ ̸ ϹѶffi

(IIT(SGO ѳѷffi ϹѶѵ ϹѵѲ ̸ ѲѴѴ

9RM1SP ѳѷffi ѲϹffi Ϲ΃ѳ ̸ Ѳffi

+SPH ѳѷffi ѵffiϹ ѵѲѴ ̸ ѳ΃

:MRE ѳѷffi ѵѶ΃ ѵѴѷ ̸ ѵѳ

%*ѳ ѲѷϹѶ GYXSJJ

TSGOIX WTIGM͸IH ѳѷffi ѵѲѱ ffiѶffi ̸ Ѷѵ΃

2YGPIMG %GMHW 4VSXIMR62% M0((8 6SWI88%*SPHѲ2% Ѳѵ ϹѶѷ Ϲѵ ̸ ѲѳѲ

%*ѳ Ѳѵ Ѳѵѳ Ѳffiѵ ̸ ѵϹѶ

4VSXIMRHW(2% M0((8 6SWI88%*SPHѲ2% ѳffi Ѳffiѳ Ѳѷ΃ ̸ ѳ΃ѵ

%*ѳ ѳffi ѳffi ѵѴ ̸ ΃Ϲ΃

'%74 Ϲѵ 62% 62% 0((8 6SWI88%*SPHѲ2% ffi ѳѵѵ Ѳffiѳ ̸ Ѵѳffi

%*ѳ ffi Ѵ΃ѳ ѴϹ΃ ̸ ѵѵѲ

%-GLIQ]C62%Ѳ LEW LYQER MRTYX

ffi ѴѳѴ Ѵѵѳ ̸ ѲѴ

62%TSPMW LEW LYQER MRTYX

ffi ѵѷѵ ѴѵѲ ̸ ѵѵffi

'LIR LEW LYQER MRTYX

ffi ѴѶffi Ѵѷ΃ ̸ ѵffiѵ

/MLEVEPEF ffi ѴѷѶ ѳѵϹ ̸ ѵѴѳ

9PXVE*SPH ffi ѳ΃ffi ѳѲѵ ̸ Ѵѵѷ

'SZEPIRX 1SH &SRHIH PMKERHW 617( Ѳ ˆ %*ѳ  ΃ffiѴ ffiѳ ̸ ffiѲ

+P]GSW]PEXMSR 617( Ѳ ˆ LMKLUYEPMX] WMRKPIVIWMHYI %*ѳ Ѳffi ΃ѱϸ ѵѳϹ ̸ ffiѵ΃

EPPUYEPMX] WMRKPIVIWMHYI %*ѳ Ϲ΃ ѳѶ Ѵѷѷ ̸ ѵѲϹ

EPPUYEPMX] QYPXMVIWMHYI %*ѳ ϹѳϹ ѳѱѳ ѳѵѴ ̸ ѴѶѳ

1SHM͸IH VIWMHYIW 617( Ѳ ˆ %*ѳ ϹѵѴ Ѵѵѵ ѵѲѴ ̸ ΃ѷ

1SHM͸IH TVSXIMR VIWMHYIW 617( Ѳ ˆ %*ѳ Ѵѷ ѴϸѶ ѳѷ ̸ ѵ

1SHM͸IH (2% VIWMHYIW 617( Ѳ ˆ %*ѳ ѶϹ ffi ѵѶѷ ̸ ΃Ѷ

1SHM͸IH 62% VIWMHYIW 617( Ѳ ˆ %*ѳ Ѳѳ ѳѶѵ ѲѳѴ ̸ ѵѶѶ

4VSXIMRW %PP 4VSXIMR4VSXIMR HSGOU \" ѷѲѳ %*1 Ѳѳ ϹѷѴ ΃ѵ Ѵ΃ ̸ ΃ѷϹ

%*ѳ ϹѷѴ ΃ ΃Ѵѷ ̸ ΃ffiѶ

4VSXIMR%RXMFSH] HSGOU \" ѷѲѳ %*1 Ѳѳ ѵ ѲѶ ϹѶ ̸ ѴѷѴ

%*ѳ ѵ ѱѵ ѵϹѴ ̸ ΃ѳѵ

1SRSQIVW 0((8 %*1 Ѳѳ ѳѳffi ffiѵѵ ffiѴ΃ ̸ ffiϹ

%*ѳ ѳѳffi ffiѵ ffiѲ ̸ ffi΃

Extended Data Table 1

ACCELERATED ARTICLE PREVIEW